Information Processing: ARTIFICIAL INTELLIGENCE
THE BRAIN BEHIND THE BRAIN
Imagine a computer with as much brainpower as all the people who have ever lived. Then add the thinking power of all the humans yet to be born before the sun dims and fades into the blackness of space some 5 billion years from now. That's a rough idea of the brain that Hugo de Garis dreams of creating.
Its neurons would be sculpted from the ultra-tiny quantum transistors that semiconductor researchers have recently developed. With these, de Garis envisions a silicon brain that packs as many neurons as a billion-billion human brains.
It's no pipe dream. De Garis has won over the Advanced Telecom Research Laboratory (ATR), which is backed by the Japanese government and Nippon Telegraph & Telephone Corp. In 1993, ATR established the Brain Builder Group in Kyoto, where de Garis has settled in for a very long-term project. "I will probably die here," says de Garis, who was born Down Under 48 years ago and earned a PhD in artificial intelligence from the Free University of Brussels.
DESIGNING. He figures it may take a decade or two to produce a silicon brain with a billion neurons, or 1% as many as people carry in their skulls. But by then, de Garis will probably have found the key to artificial brains of unimaginable complexity: a process of electronic evolution. It will not only coax silicon circuits into giving birth to innate intelligence but imbue them with the power to design themselves--to control their own destiny by spawning new generations of ever improving brains at electronic speeds.
The three-person Brain Builder Group is part of a $190 million, nine-year project aimed at creating computer chips that mimic the human brain and nervous system. Other ATR researchers are working to reconstruct vision, hearing, and other senses in silicon circuits. The unifying goal is to create a new computer architecture--a "hardware habitation"--that will enable smart machines to grow and evolve by Darwinian principles of survival of the fittest. In addition to ATR, a handful of other labs in Europe and the U.S. are working along similar lines, christened "evolvable hardware."
Already, de Garis is using a set of 11,000 artificial-evolution rules to guide the growth of simulated neurons and synapses on a two-dimensional plane. "But this is just a toy," de Garis admits, because these circuits are limited to a few connections each. For comparison, some human neurons connect to 100,000 other brain cells. "So the real complexity will come in three dimensions," says de Garis. Even with his 2-D plaything, though, he has evolved circuits that function like a primitive retina, among other simple feats.
Now, de Garis is working on a new set of 80,000 rules that will run on a special breed of computer developed for artificial-intelligence research at Massachusetts Institute of Technology. Called a cellular automata machine, this $40,000 computer packs the punch needed to grow 3-D circuits. De Garis expects to simulate a kitten in a year or so.
Even within the small evolvable-hardware fraternity, de Garis is the odd man out. Skeptics say he puts too much trust in computer simulations and needs to run periodic reality checks. Inman Harvey, a mathematician who helped launch the evolvable-hardware group at Sussex University, asserts that simulations alone can't provide an environment rich enough to nurture the superbrain that de Garis wants to create. "Look out the window, and you've got an incredible amount of information coming at you at the speed of light," says Harvey. "That's why I and some others doubt his optimistic projections."
De Garis remains adamant. Using real-world experience to hone survival traits in lumbering electromechanical robots, he retorts, would severely hamper progress. Evolving a superbrain that way might take centuries, not decades. "If you don't simulate," he insists, "you don't get enough speed."
Slowing things down might not be such a bad idea, since it's doubtful that a couple of decades will allow sufficient time for people to adjust to the notion of ranking a poor second on the intelligence scale. Once it becomes clear that silicon evolution could unleash a new entity that might quickly advance beyond human control, de Garis and others envision fierce ideological debates over whether the building of such alien brains should be permitted. Even de Garis has reservations: "I'm worried, too, because I don't want to be swatted like a fly."
Despite the risk that his dream machine might turn into a science-fiction nightmare, de Garis believes it will be built, eventually--if not by scientists bent on unwrapping the last of nature's impenetrable mysteries, then by corporate researchers out to one-up the competition. The potential payoff from silicon brains only a little smarter than people could be enormous. And by then, the point of no return would probably have been passed.By Larry Holyoke in Kyoto, with Otis Port in New York